Databricks claims DBRX sets ‘a new standard’ for open-source LLMs

Databricks has launched DBRX, an open-source large language model with 132 billion parameters. It outperforms established options such as GPT-3.5, LLaMA 2 70B, Mixtral, Grok-1, and even Anthropic’s closed-source model Claude on specific benchmarks. DBRX excels in language understanding, programming, and maths tasks, showcasing state-of-the-art performance in coding tasks and demonstrating competitive or superior performance to GPT-3.5 across multiple benchmarks.

The state-of-the-art DBRX model is 2x faster at inference than LLaMA 2 70B and was trained to be more compute-efficient. Databricks co-founder and CEO, Ali Ghodsi, highlighted the platform’s capacity for custom reasoning based on enterprise data. The model was trained on 12 trillion tokens of curated text and code data and can be fine-tuned using APIs or the company’s tools. Databricks aims to address data security and privacy concerns, and partners such as Accenture, Block, Nasdaq, Prosus, Replit, and Zoom have praised DBRX’s potential. NASDAQ’s Head of AI and Data Services expressed excitement about the release of DBRX, emphasizing its strong performance and favorable serving economics.